Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,986 Bytes
fb5e185 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 |
import math
from dataclasses import dataclass
from typing import List, Optional, Union
import numpy as np
import torch
import torch.nn.functional as F
# import trimesh
from PIL import Image
from torch import BoolTensor, FloatTensor
LIST_TYPE = Union[list, np.ndarray, torch.Tensor]
def list_to_pt(
x: LIST_TYPE, dtype: Optional[torch.dtype] = None, device: Optional[str] = None
) -> torch.Tensor:
if isinstance(x, list) or isinstance(x, np.ndarray):
return torch.tensor(x, dtype=dtype, device=device)
return x.to(dtype=dtype)
def get_c2w(
elevation_deg: LIST_TYPE,
distance: LIST_TYPE,
azimuth_deg: Optional[LIST_TYPE],
num_views: Optional[int] = 1,
device: Optional[str] = None,
) -> torch.FloatTensor:
if azimuth_deg is None:
assert (
num_views is not None
), "num_views must be provided if azimuth_deg is None."
azimuth_deg = torch.linspace(
0, 360, num_views + 1, dtype=torch.float32, device=device
)[:-1]
else:
num_views = len(azimuth_deg)
azimuth_deg = list_to_pt(azimuth_deg, dtype=torch.float32, device=device)
elevation_deg = list_to_pt(elevation_deg, dtype=torch.float32, device=device)
camera_distances = list_to_pt(distance, dtype=torch.float32, device=device)
elevation = elevation_deg * math.pi / 180
azimuth = azimuth_deg * math.pi / 180
camera_positions = torch.stack(
[
camera_distances * torch.cos(elevation) * torch.cos(azimuth),
camera_distances * torch.cos(elevation) * torch.sin(azimuth),
camera_distances * torch.sin(elevation),
],
dim=-1,
)
center = torch.zeros_like(camera_positions)
up = torch.tensor([0, 0, 1], dtype=torch.float32, device=device)[None, :].repeat(
num_views, 1
)
lookat = F.normalize(center - camera_positions, dim=-1)
right = F.normalize(torch.cross(lookat, up, dim=-1), dim=-1)
up = F.normalize(torch.cross(right, lookat, dim=-1), dim=-1)
c2w3x4 = torch.cat(
[torch.stack([right, up, -lookat], dim=-1), camera_positions[:, :, None]],
dim=-1,
)
c2w = torch.cat([c2w3x4, torch.zeros_like(c2w3x4[:, :1])], dim=1)
c2w[:, 3, 3] = 1.0
return c2w
def get_projection_matrix(
fovy_deg: LIST_TYPE,
aspect_wh: float = 1.0,
near: float = 0.1,
far: float = 100.0,
device: Optional[str] = None,
) -> torch.FloatTensor:
fovy_deg = list_to_pt(fovy_deg, dtype=torch.float32, device=device)
batch_size = fovy_deg.shape[0]
fovy = fovy_deg * math.pi / 180
tan_half_fovy = torch.tan(fovy / 2)
projection_matrix = torch.zeros(
batch_size, 4, 4, dtype=torch.float32, device=device
)
projection_matrix[:, 0, 0] = 1 / (aspect_wh * tan_half_fovy)
projection_matrix[:, 1, 1] = -1 / tan_half_fovy
projection_matrix[:, 2, 2] = -(far + near) / (far - near)
projection_matrix[:, 2, 3] = -2 * far * near / (far - near)
projection_matrix[:, 3, 2] = -1
return projection_matrix
def get_orthogonal_projection_matrix(
batch_size: int,
left: float,
right: float,
bottom: float,
top: float,
near: float = 0.1,
far: float = 100.0,
device: Optional[str] = None,
) -> torch.FloatTensor:
projection_matrix = torch.zeros(
batch_size, 4, 4, dtype=torch.float32, device=device
)
projection_matrix[:, 0, 0] = 2 / (right - left)
projection_matrix[:, 1, 1] = -2 / (top - bottom)
projection_matrix[:, 2, 2] = -2 / (far - near)
projection_matrix[:, 0, 3] = -(right + left) / (right - left)
projection_matrix[:, 1, 3] = -(top + bottom) / (top - bottom)
projection_matrix[:, 2, 3] = -(far + near) / (far - near)
projection_matrix[:, 3, 3] = 1
return projection_matrix
@dataclass
class Camera:
c2w: Optional[torch.FloatTensor]
w2c: torch.FloatTensor
proj_mtx: torch.FloatTensor
mvp_mtx: torch.FloatTensor
cam_pos: Optional[torch.FloatTensor]
def __getitem__(self, index):
if isinstance(index, int):
sl = slice(index, index + 1)
elif isinstance(index, slice):
sl = index
else:
raise NotImplementedError
return Camera(
c2w=self.c2w[sl] if self.c2w is not None else None,
w2c=self.w2c[sl],
proj_mtx=self.proj_mtx[sl],
mvp_mtx=self.mvp_mtx[sl],
cam_pos=self.cam_pos[sl] if self.cam_pos is not None else None,
)
def to(self, device: Optional[str] = None):
if self.c2w is not None:
self.c2w = self.c2w.to(device)
self.w2c = self.w2c.to(device)
self.proj_mtx = self.proj_mtx.to(device)
self.mvp_mtx = self.mvp_mtx.to(device)
if self.cam_pos is not None:
self.cam_pos = self.cam_pos.to(device)
def __len__(self):
return self.c2w.shape[0]
def get_camera(
elevation_deg: Optional[LIST_TYPE] = None,
distance: Optional[LIST_TYPE] = None,
fovy_deg: Optional[LIST_TYPE] = None,
azimuth_deg: Optional[LIST_TYPE] = None,
num_views: Optional[int] = 1,
c2w: Optional[torch.FloatTensor] = None,
w2c: Optional[torch.FloatTensor] = None,
proj_mtx: Optional[torch.FloatTensor] = None,
aspect_wh: float = 1.0,
near: float = 0.1,
far: float = 100.0,
device: Optional[str] = None,
):
if w2c is None:
if c2w is None:
c2w = get_c2w(elevation_deg, distance, azimuth_deg, num_views, device)
camera_positions = c2w[:, :3, 3]
w2c = torch.linalg.inv(c2w)
else:
camera_positions = None
c2w = None
if proj_mtx is None:
proj_mtx = get_projection_matrix(
fovy_deg, aspect_wh=aspect_wh, near=near, far=far, device=device
)
mvp_mtx = proj_mtx @ w2c
return Camera(
c2w=c2w, w2c=w2c, proj_mtx=proj_mtx, mvp_mtx=mvp_mtx, cam_pos=camera_positions
)
def get_orthogonal_camera(
elevation_deg: LIST_TYPE,
distance: LIST_TYPE,
left: float,
right: float,
bottom: float,
top: float,
azimuth_deg: Optional[LIST_TYPE] = None,
num_views: Optional[int] = 1,
near: float = 0.1,
far: float = 100.0,
device: Optional[str] = None,
):
c2w = get_c2w(elevation_deg, distance, azimuth_deg, num_views, device)
camera_positions = c2w[:, :3, 3]
w2c = torch.linalg.inv(c2w)
proj_mtx = get_orthogonal_projection_matrix(
batch_size=c2w.shape[0],
left=left,
right=right,
bottom=bottom,
top=top,
near=near,
far=far,
device=device,
)
mvp_mtx = proj_mtx @ w2c
return Camera(
c2w=c2w, w2c=w2c, proj_mtx=proj_mtx, mvp_mtx=mvp_mtx, cam_pos=camera_positions
)
|